Monolith Research
The sentiment work remains in the file because failed replications are useful. The pipeline built daily macro-event features from GDELT, enriched the headline stream, and tested both tree-based and linear models against EUR/USD and USD/JPY. Thirty-six features were constructed: tone, article volume, Goldstein measures, rolling sentiment statistics, lags, and market-state variables.
The best classification reading was still too close to nothing. AUC stayed near 0.50, with the strongest published run in the local audit only around 0.510. Sharpe was negative across the live-data replications. FinBERT-style enrichment did not rescue the result; in the recorded runs it made the trading output worse rather than better.
That does not make the market-tone system useless. It changes its job. The research branch should not size trades, select entries, or promote confidence. It can remain as a warning layer: a way to mark crowded news days, macro-event density, and moments where the desk should slow down before trusting a clean-looking signal.
The best statistic in this note is not a return. It is the refusal to deploy. A coin-flip model, written down honestly, is cheaper than a beautiful one allowed to trade.